VELVET VICE MiniMax H3 — AUTO Text Encoder Loader

MiniMax H3 doesn't condition on a CLIP text encoder the way a Stable Diffusion checkpoint does. Its prompts go through a Qwen3-VL encoder - and that encoder can arrive on your disk two very different ways: as native safetensors, or as a GGUF quantized file to save VRAM. The VelvetViceMiniMaxH3CLIPLoader node is the single AUTO selector that loads whichever one you have, and it's the closest thing this pack has to a boring drop-in replacement for ComfyUI's stock CLIPLoader.

The node's clip_name input is literally a dropdown with one option: AUTO. That's not a placeholder - it's the whole design. Instead of making you pick a file from a folder and pray it matches what the model expects, the loader scans ComfyUI's known text-encoder folders (including legacy aliases and subfolders), finds the H3-compatible Qwen3-VL file, and loads it.

How the auto-matching works

Under the hood the node checks the resolved filename's extension. If it ends in .gguf, it looks up ComfyUI-GGUF's CLIPLoaderGGUF and calls it with the minimax type - so you don't need a separate GGUF clip loader node. If it's native, it falls back to ComfyUI's own CLIPLoader with the same minimax type. "AUTO matches compatible filenames across current and legacy ComfyUI folder aliases, including subfolders" - which is the pack's way of handling the fact that one ComfyUI release calls the folder text_encoders and another calls it clip, with GGUF plugins adding clip_gguf on top. The catalog de-duplicates across all of them, so the dropdown stays stable regardless of load order.

There's a device selector too: default or cpu. Picking cpu offloads the Qwen3-VL encoder to system RAM. Given H3's diffusion model is enormous and the encoder is comparatively small, running the encoder on CPU is a legitimately smart VRAM trade if you're squeezing onto a card.

What it gives you

• clip_name (AUTO) - the auto-resolving selector.
• device (default / cpu) - where the encoder runs.

Output: a single CLIP value to feed your sampler's positive/negative conditioning.

Install and gotchas

Part of the velvet-vice-minimax-h3 pack:

cd ComfyUI/custom_nodes
git clone https://github.com/Velvet-Vice/velvet-vice-minimax-h3

then restart. If you plan to use the GGUF path, you need ComfyUI-GGUF installed - the node is explicit about this and raises a clear "ComfyUI-GGUF / CLIPLoaderGGUF is not installed" error if you select a GGUF encoder without it. Native safetensors needs no extra plugin.

Two things worth knowing before you treat this as a black box. First, AUTO only works when exactly the expected H3 file is present; if you have multiple Qwen3-VL variants in your folders the resolution logic picks the compatible one, and if none match you'll see an empty or unresolved load rather than a random encoder. Second, whatever type string ComfyUI's H3 sampler expects (minimax), both code paths here pass it - so don't hand-roll a stock CLIPLoader with the wrong type when this node exists for exactly that reason.